Highlights:

  • The models developed by the company have the ability to transcribe speech, categorize the text into chapters, and link each section of the transcript to the appropriate speaker.
  • Some thousands of organizations utilize the company’s models for processing the voice data of their applications.

AssemblyAI Inc., a startup that creates AI models tailored to handle audio recordings, has raised USD 50 million in late-stage capital to assist its expansion.

The company made the Series C funding announcement recently. Insight Partners, Co-chief Executive of Salesforce Inc. Keith Block, former GitHub CEO Nat Friedman, and several additional supporters joined Accel, the lead investor. With this round, AssemblyAI has received USD 115 million in outside capital, most of which has been raised since the year’s beginning.

AssemblyAI is a 2017 startup that provides cloud-based AI models for audio files, such as call center records. The models developed by the company have the ability to transcribe speech, categorize the text into chapters, and link each section of the transcript to the appropriate speaker.

Additionally, AssemblyAI assists developers with content analysis of their audio files. The business’s AI models may automatically create summaries by extracting exciting information from transcripts, such as references of competitors. It also provides a neural network designed to identify and eliminate sensitive information, including client credit card numbers.

Conformer-2, AssemblyAI’s most recent AI model, debuted in July. Approximately 1.1 million hours of voice data were used to train it. According to the business, Conformer-2 interprets audio files around 50% more accurately than its previous-generation neural networks.

AssemblyAI offers a LeMURE development framework in addition to cloud-based AI models. Software teams typically have to use multiple distinct development tools to integrate a neural network into an application, which makes their work more hectic. Software teams can use LeMURE instead of those tools, according to AssemblyAI, to expedite the development process.

Some thousands of organizations utilize the company’s models for processing the voice data of their applications. Dylan Fox, Founder and CEO of AssemblyAI, reported, “We’re now regularly serving over 25 million inference calls and processing over 10 terabytes of voice data, every day through our API for our customers. And, with 10,000-plus new organizations signing up for our API every month, we’re just scratching the surface of the new voice-powered AI applications we’ll see enter the market over the next year.”

The company plans to intensify its client acquisition initiatives and hire more staff with the capital raised from this fresh fundraising round. Furthermore, AssemblyAI intends to acquire additional computational infrastructure tailored for AI training. The business will use this infrastructure to diversify its neural network offerings.

AssemblyAI is now developing a new flagship AI model called Universal, which is anticipated to perform better than its most recent Conformer-2 system in several domains. The dataset the company uses to train Universal is around ten times bigger than the one utilized to create Conformer-2. Google Cloud instances powered by the search giant’s in-house TPU chips—optimized for AI workloads—are used for the training.

About six months ago, AssemblyAI started working on Universal. According to Fox’s recent statement, the model will be made accessible to users very soon.